Dec. 19, 2024
Pulmonary hypertension (PH) is a progressive and life-threatening disease associated with significant morbidity and mortality. The symptoms of pulmonary hypertension are nonspecific, so delays in diagnosis are common and associated with worse outcomes.
A precise, noninvasive screening tool for PH is needed to improve early detection and treatment. Machine learning can facilitate the detection of subtle, disease-specific abnormalities on routine tests, such as electrocardiograms (ECGs), which are commonly performed in the diagnostic evaluation of PH and presenting symptoms, such as dyspnea.
"In a retrospective cohort study published in the July 2024 issue of European Respiratory Journal, an electrocardiogram-based artificial intelligence algorithm was applied to standard 12-lead electrocardiograms to develop an early detection tool for PH," says Hilary M. DuBrock, M.D., a pulmonologist and critical care specialist at Mayo Clinic in Rochester, Minnesota.
Each standard 12-lead electrocardiogram was paired with clinical data to define patients as PH-likely, defined as a mean pulmonary arterial pressure (mPAP) > 20 mm Hg at rest by right heart catheterization (RHC) or tricuspid regurgitation velocity > 3.4 m/s on echocardiogram if RHC was unavailable (n = 39,823) or PH-unlikely, defined as a mPAP ≤ 20 mm Hg or echocardiographic tricuspid regurgitation velocity ≤ 2.8 m/s if RHC was unavailable (n = 219,404).
Patients were randomly split into training (48%), validation (12%) and test sets (40%) for developing, optimizing and testing the model using convolutional neural networks. ECGs taken within one month of PH diagnosis (diagnostic test sets) were used to train the model. Performance was tested on ECGs from both Mayo Clinic and an external validation cohort at Vanderbilt University Medical Center at diagnosis and up to five years prior to PH diagnosis.
The final model was able to detect PH within one month of diagnosis with an area under the curve (AUC), sensitivity and specificity of 0.92, 85.5% and 83.8% at Mayo Clinic and 0.88, 82.3% and 77.0% at Vanderbilt University, respectively. Importantly, the model was also able to detect PH at Mayo Clinic and Vanderbilt University with an AUC of 0.86 and 0.81 at 6 to 18 months prior to diagnosis and 0.79 and 0.73 up to five years before diagnosis. The algorithm has been granted Food and Drug Administration breakthrough designation and is currently being evaluated in a multicenter study.
In summary, the PH early detection algorithm was able to detect PH at the time of clinical diagnosis and 6 to 18 months prior to diagnosis at two independent medical centers. It has the potential to accelerate diagnosis and treatment of PH which could ultimately lead to improved patient outcomes.
For more information
DuBrock HM, et al. An electrocardiogram-based AI algorithm for early detection of pulmonary hypertension.
European Respiratory Journal. 2024;64:2400192.
Refer a patient to Mayo Clinic.